Update repositories, reenable ROCm on selfhosted runner, add packages (…

…#24) CI: * Rename actions * Create actions generator * Configure mirrorlist before building * Use work dir inside container * Remove container when stopped * Fix CI exec conditions Misc: * Update repositories * Add useful packages * Update README.md * Add vm shutdown logic
sasha0552 · Feb 14, 2024 · cf09f61 · cf09f61
1 parent c43bcc8
commit cf09f61
Show file tree

Hide file tree

Showing 16 changed files with 248 additions and 23 deletions.
diff --git a/.ci/actions.py b/.ci/actions.py
@@ -0,0 +1,28 @@
+#!/usr/bin/env python3
+
+import jinja2
+
+def main():
+  for filename in [ "gh-build-iso", "sh-build-iso" ]:
+    for type in [ "cuda", "rocm" ]:
+      # read input file
+      with open(f".ci/template/{filename}.yml.jinja2", "r") as file:
+        template = jinja2.Template(file.read())
+
+      # render template
+      rendered = template.render(type=type)
+
+      # FIXME: skip if hosted and rocm
+      if filename == "gh-build-iso" and type == "rocm":
+        continue
+
+      # FIXME: skip if selfhosted and cuda
+      if filename == "sh-build-iso" and type == "cuda":
+        continue
+
+      # write output file
+      with open(f".github/workflows/{filename}-{type}.yml", "w") as file:
+        file.write(rendered)
+
+if __name__ == "__main__":
+  main()
diff --git a/.ci/dependencies.sh b/.ci/dependencies.sh
@@ -17,4 +17,7 @@ pacman-key --init
 pacman-key --populate
 
 # Install required packages
-pacman --sync --noconfirm --needed archiso patch python python-jinja rdfind
+pacman --sync --noconfirm --needed archiso patch python python-jinja rdfind reflector
+
+# Configure mirrorlist
+reflector --latest 5 --sort rate --save /etc/pacman.d/mirrorlist
diff --git a/...hub/workflows/build-iso-rocm.yml.disabled → .ci/template/gh-build-iso.yml.jinja2 b/...hub/workflows/build-iso-rocm.yml.disabled → .ci/template/gh-build-iso.yml.jinja2
@@ -1,8 +1,13 @@
-name: Build ISO (ROCm)
+name: Build ISO on hosted runner ({{ "CUDA" if type == "cuda" else "ROCm" }})
 
 on:
-  - push
-  - pull_request
+  push:
+    branches:
+    - main
+
+  pull_request:
+    branches:
+    - main
 
 jobs:
   build:
@@ -23,7 +28,9 @@ jobs:
         uses: addnab/docker-run-action@v3
         with:
           image: archlinux:latest
+{%- raw %}
           options: --privileged --volume ${{ github.workspace }}:/workspace
+{%- endraw %}
           run: |
             # Exit on error
             set -eu
@@ -36,12 +43,12 @@ jobs:
               # Patch mkarchiso
               .ci/mkarchiso.sh
 
-              # Configure to use ROCm
-              .ci/configure.py rocm
+              # Configure to use {{ "CUDA" if type == "cuda" else "ROCm" }}
+              .ci/configure.py {{ type }}
             popd
 
             # Build image
-            mkarchiso -v -m iso -w /workspace/work -o /workspace/out /workspace
+            mkarchiso -v -m iso -w /_work -o /workspace/out /workspace
 
       - name: Create summary
         run: |

diff --git a/.ci/template/sh-build-iso.yml.jinja2 b/.ci/template/sh-build-iso.yml.jinja2
@@ -0,0 +1,90 @@
+name: Build ISO on selfhosted runner ({{ "CUDA" if type == "cuda" else "ROCm" }})
+
+on:
+  push:
+    branches:
+    - main
+
+  pull_request:
+    branches:
+    - main
+
+jobs:
+  vm-start:
+    runs-on: ubuntu-latest
+
+    steps:
+      - name: Start virtual machine
+        run: |
+          # Do nothing for now
+          exit 0
+
+  build:
+    needs: vm-start
+    runs-on: self-hosted
+
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@v4
+        with:
+          submodules: recursive
+
+      - name: Build image
+        uses: addnab/docker-run-action@v3
+        with:
+          image: archlinux:latest
+{%- raw %}
+          options: --privileged --rm --volume ${{ github.workspace }}:/workspace
+{%- endraw %}
+          run: |
+            # Exit on error
+            set -eu
+
+            # Enter project directory
+            pushd /workspace
+              # Install dependencies
+              .ci/dependencies.sh
+
+              # Patch mkarchiso
+              .ci/mkarchiso.sh
+
+              # Configure to use {{ "CUDA" if type == "cuda" else "ROCm" }}
+              .ci/configure.py {{ type }}
+            popd
+
+            # Build image
+            mkarchiso -v -m iso -w /_work -o /workspace/out /workspace
+
+      - name: Create summary
+        run: |
+          # Exit on error
+          set -eu
+
+          # Print checksums to summary
+          sha256sum out/* > "$GITHUB_STEP_SUMMARY"
+
+      - name: Upload artifacts
+        uses: actions/upload-artifact@v4
+        with:
+          name: archiso-output
+          path: out/
+
+  vm-shutdown:
+    needs: build
+    runs-on: self-hosted
+
+    steps:
+      - name: Stop virtual machine
+        run: |
+          # Shutdown in 1 minute
+          sudo shutdown 1
+
+  vm-shutdown-wait:
+    needs: vm-shutdown
+    runs-on: ubuntu-latest
+
+    steps:
+      - name: Wait until virtual machine is stopped
+        run: |
+          # Wait for 2 minutes
+          sleep 120
diff --git a/.github/workflows/build-iso-cuda.yml → .github/workflows/gh-build-iso-cuda.yml b/.github/workflows/build-iso-cuda.yml → .github/workflows/gh-build-iso-cuda.yml
@@ -1,8 +1,13 @@
-name: Build ISO (CUDA)
+name: Build ISO on hosted runner (CUDA)
 
 on:
-  - push
-  - pull_request
+  push:
+    branches:
+    - main
+
+  pull_request:
+    branches:
+    - main
 
 jobs:
   build:
@@ -41,7 +46,7 @@ jobs:
             popd
 
             # Build image
-            mkarchiso -v -m iso -w /workspace/work -o /workspace/out /workspace
+            mkarchiso -v -m iso -w /_work -o /workspace/out /workspace
 
       - name: Create summary
         run: |
@@ -55,4 +60,4 @@ jobs:
         uses: actions/upload-artifact@v4
         with:
           name: archiso-output
-          path: out/
+          path: out/
diff --git a/.github/workflows/sh-build-iso-rocm.yml b/.github/workflows/sh-build-iso-rocm.yml
@@ -0,0 +1,88 @@
+name: Build ISO on selfhosted runner (ROCm)
+
+on:
+  push:
+    branches:
+    - main
+
+  pull_request:
+    branches:
+    - main
+
+jobs:
+  vm-start:
+    runs-on: ubuntu-latest
+
+    steps:
+      - name: Start virtual machine
+        run: |
+          # Do nothing for now
+          exit 0
+
+  build:
+    needs: vm-start
+    runs-on: self-hosted
+
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@v4
+        with:
+          submodules: recursive
+
+      - name: Build image
+        uses: addnab/docker-run-action@v3
+        with:
+          image: archlinux:latest
+          options: --privileged --rm --volume ${{ github.workspace }}:/workspace
+          run: |
+            # Exit on error
+            set -eu
+
+            # Enter project directory
+            pushd /workspace
+              # Install dependencies
+              .ci/dependencies.sh
+
+              # Patch mkarchiso
+              .ci/mkarchiso.sh
+
+              # Configure to use ROCm
+              .ci/configure.py rocm
+            popd
+
+            # Build image
+            mkarchiso -v -m iso -w /_work -o /workspace/out /workspace
+
+      - name: Create summary
+        run: |
+          # Exit on error
+          set -eu
+
+          # Print checksums to summary
+          sha256sum out/* > "$GITHUB_STEP_SUMMARY"
+
+      - name: Upload artifacts
+        uses: actions/upload-artifact@v4
+        with:
+          name: archiso-output
+          path: out/
+
+  vm-shutdown:
+    needs: build
+    runs-on: self-hosted
+
+    steps:
+      - name: Stop virtual machine
+        run: |
+          # Shutdown in 1 minute
+          sudo shutdown 1
+
+  vm-shutdown-wait:
+    needs: vm-shutdown
+    runs-on: ubuntu-latest
+
+    steps:
+      - name: Wait until virtual machine is stopped
+        run: |
+          # Wait for 2 minutes
+          sleep 120
diff --git a/README.md b/README.md
@@ -22,7 +22,7 @@ If you would like to see another AI-related project included in ToriLinux, pleas
 
 To use ToriLinux:
 1. Install [Ventoy](https://ventoy.net/en/doc_start.html) on a USB drive.
-2. Download the latest ISO from workflows ([NVIDIA](https://github.com/sasha0552/ToriLinux/actions/workflows/build-iso-cuda.yml?query=branch%3Amain) / [AMD](https://github.com/sasha0552/ToriLinux/actions/workflows/build-iso-rocm.yml?query=branch%3Amain)) and copy it to the USB drive.
+2. Download the latest ISO from workflows ([NVIDIA](https://github.com/sasha0552/ToriLinux/actions/workflows/gh-build-iso-cuda?query=branch%3Amain) / [AMD](https://github.com/sasha0552/ToriLinux/actions/workflows/sh-build-iso-rocm.yml?query=branch%3Amain)) and copy it to the USB drive.
 3. Boot from the USB drive (select it as the boot device in BIOS/UEFI).
 4. Log in with the username `tori` and password `tori`. You can also use [SSH](https://en.wikipedia.org/wiki/Secure_Shell).
 
@@ -33,3 +33,5 @@ Note that you need pre-downloaded models on a local hard drive or NFS server, or
 Note that following projects is not available on ROCm version:
 * [axolotl](https://github.com/OpenAccess-AI-Collective/axolotl)
 * [vllm](https://github.com/vllm-project/vllm)
+
+The server for building the ROCm version is provided by [@Sepera-okeq](https://github.com/Sepera-okeq/).
diff --git a/airootfs/home/tori/ComfyUI b/airootfs/home/tori/ComfyUI
diff --git a/airootfs/home/tori/automatic b/airootfs/home/tori/automatic
diff --git a/airootfs/home/tori/axolotl b/airootfs/home/tori/axolotl
diff --git a/airootfs/home/tori/koboldcpp b/airootfs/home/tori/koboldcpp
diff --git a/airootfs/home/tori/llama.cpp b/airootfs/home/tori/llama.cpp
diff --git a/airootfs/home/tori/vllm b/airootfs/home/tori/vllm
diff --git a/airootfs/root/customize_airootfs/patches/0100-llamacpp-enable-prompt-cache.patch b/airootfs/root/customize_airootfs/patches/0100-llamacpp-enable-prompt-cache.patch
@@ -1,17 +1,17 @@
 --- a/examples/server/oai.hpp
 +++ b/examples/server/oai.hpp
-@@ -31,7 +31,7 @@ inline static json oaicompat_completion_params_parse(
+@@ -35,7 +35,7 @@ inline static json oaicompat_completion_params_parse(
      llama_sampling_params default_sparams;
      llama_params["model"]             = json_value(body, "model", std::string("unknown"));
-     llama_params["prompt"]            = format_chatml(body["messages"]); // OpenAI 'messages' to llama.cpp 'prompt'
+     llama_params["prompt"]            = formatted_prompt;
 -    llama_params["cache_prompt"]      = json_value(body, "cache_prompt", false);
 +    llama_params["cache_prompt"]      = json_value(body, "cache_prompt", true);
      llama_params["temperature"]       = json_value(body, "temperature", 0.0);
      llama_params["top_k"]             = json_value(body, "top_k", default_sparams.top_k);
      llama_params["top_p"]             = json_value(body, "top_p", 1.0);
 --- a/examples/server/server.cpp
 +++ b/examples/server/server.cpp
-@@ -525,7 +525,7 @@ struct llama_server_context
+@@ -526,7 +526,7 @@ struct llama_server_context
          }
 
          slot->params.stream             = json_value(data, "stream",            false);

diff --git a/airootfs/root/customize_airootfs/patches/0100-vllm-enable-other-archs.patch b/airootfs/root/customize_airootfs/patches/0100-vllm-enable-other-archs.patch
@@ -6,7 +6,7 @@
  # Supported NVIDIA GPU architectures.
 -NVIDIA_SUPPORTED_ARCHS = {"7.0", "7.5", "8.0", "8.6", "8.9", "9.0"}
 +NVIDIA_SUPPORTED_ARCHS = {"6.0", "6.1", "7.0", "7.5", "8.0", "8.6", "8.9", "9.0"}
- ROCM_SUPPORTED_ARCHS = {"gfx90a", "gfx942"}
+ ROCM_SUPPORTED_ARCHS = {"gfx90a", "gfx942", "gfx1100"}
  # SUPPORTED_ARCHS = NVIDIA_SUPPORTED_ARCHS.union(ROCM_SUPPORTED_ARCHS)
 
 @@ -222,9 +222,9 @@ if _is_cuda() and not compute_capabilities:

diff --git a/packages.x86_64.jinja2 b/packages.x86_64.jinja2
@@ -20,7 +20,9 @@ mkinitcpio-nfs-utils
 nano
 ncdu
 neofetch
+net-tools
 nfs-utils
+ntfs-3g
 nvtop
 openssh
 progress
+1 −1		comfy/controlnet.py
+4 −0		comfy/model_base.py
+3 −2		comfy/samplers.py
+6 −3		comfy/sd.py
+25 −0		comfy_extras/nodes_cond.py
+20 −0		comfy_extras/nodes_images.py
+1 −0		nodes.py
+1 −0		web/extensions/core/widgetInputs.js
+6 −1		web/style.css
+61 −2		convert-hf-to-gguf.py
+2 −2		examples/llava/README.md
+86 −47		ggml-cuda.cu
+5 −3		ggml-sycl.cpp
+4 −3		ggml-vulkan.cpp
+7 −0		gpttype_adapter.cpp
+28 −17		klite.embd
+5 −2		koboldcpp.py
+8 −6		llama.cpp
+1 −0		.flake8
+1 −1		.github/workflows/python-lint.yml
+19 −5		Package.swift
+1 −0		README.md
+85 −35		common/common.cpp
+5 −2		common/common.h
+14 −21		common/sampling.cpp
+19 −1		common/sampling.h
+94 −0		convert-hf-to-gguf.py
+2 −1		convert-persimmon-to-gguf.py
+11 −1		examples/embedding/embedding.cpp
+5 −14		examples/export-lora/export-lora.cpp
+37 −108		examples/finetune/finetune.cpp
+9 −3		examples/llava/README.md
+81 −71		examples/llava/clip.cpp
+1 −1		examples/llava/convert-image-encoder-to-gguf.py
+1 −1		examples/llava/llava-surgery.py
+3 −0		examples/llava/requirements.txt
+11 −1		examples/lookup/lookup.cpp
+5 −2		examples/main/main.cpp
+1 −1		examples/server/README.md
+6 −2		examples/server/oai.hpp
+49 −13		examples/server/server.cpp
+30 −0		examples/server/utils.hpp
+32 −80		examples/train-text-from-scratch/train-text-from-scratch.cpp
+3 −3		flake.lock
+563 −490		ggml-alloc.c
+39 −65		ggml-alloc.h
+250 −262		ggml-backend.c
+8 −12		ggml-backend.h
+140 −107		ggml-cuda.cu
+2 −0		ggml-metal.m
+329 −22		ggml-quants.c
+14 −14		ggml-quants.h
+14 −61		ggml-sycl.cpp
+89 −6		ggml-vulkan.cpp
+134 −60		ggml.c
+26 −6		ggml.h
+10 −1		ggml_vk_generate_shaders.py
+25 −18		gguf-py/gguf/constants.py
+6 −0		gguf-py/gguf/gguf_writer.py
+10 −3		gguf-py/gguf/tensor_mapping.py
+580 −114		llama.cpp
+1 −0		llama.h
+2 −2		pocs/vdot/q8dot.cpp
+2 −2		pocs/vdot/vdot.cpp
+12 −0		scripts/sync-ggml-am.sh
+1 −1		scripts/sync-ggml.last
+6 −0		scripts/sync-ggml.sh
+1 −0		spm-headers/ggml-alloc.h
+1 −0		spm-headers/ggml-backend.h
+1 −0		spm-headers/ggml.h
+1 −1		tests/test-quantize-fns.cpp
+1 −1		tests/test-quantize-perf.cpp
+12 −3		Dockerfile.rocm
+72 −294		csrc/quantization/awq/gemm_kernels.cu
+2 −1		docs/source/getting_started/amd-installation.rst
+1 −0		docs/source/index.rst
+52 −0		docs/source/models/lora.rst
+1 −1		setup.py
+45 −0		vllm/model_executor/layers/attention.py
+1 −1		vllm/model_executor/layers/quantization/awq.py